Grok 4.20 Beta had the lowest hallucination rate of 26 models on AA-Omniscience, 22%; gpt-oss-20B had 94%

What the chart shows

AA-Omniscience asks 6,000 questions in six domains, from law and health to software engineering, and does not penalise a model for declining to answer. Its hallucination rate measures how often models give false answers instead of acknowledging uncertainty. Grok 4.20 Beta 0305 had the lowest rate of the 26 models, 22%, followed by Claude 4.5 Haiku (26%) and MiMo-V2-Pro (30%).

At the other end, gpt-oss-20B (high) reached 94% and Gemini 3 Flash 92%. More than half of the models, 15 of 26, had rates of 82% or higher.

Key numbers

How to read it

• Scoring rewards correct answers, penalises incorrect ones and does not penalise refusals. The AI Index notes that a lower rate means a model knows more or is better at recognising when it is unsure. The scale is not comparable with Vectara’s summary test (related chart).
• Settings in brackets (high, xhigh, max, medium) are the reasoning settings tested.
• The data file and the report’s chart label one model gpt-oss-128B; OpenAI’s model is gpt-oss-120B, the name the AI Index uses in its openness chart, so we show that.